Fedora 43 Cuda 13 article ideas, followup from popular-ish article

[level1mo]

Hey folks,

I am looking to create a new article to supplement the one that I created before on setting up CUDA with fedora.

Currently, the article I have in mind is an updated one of the above but with even more BONUS sections including:

• desktop setups including: setting up gnome tweaks, hardware acceleration on firefox, and so one
• data science notebooks: setting up a jupyterlab and marimo to use with CUDA 13 including using uv for keeping everything together and latest pytorch with an example running
• Going through a C++ thrust example (maybe writing a mandlebrot version in a comment)
• setting up a local ai stack (NVFP4/MXFP4 llm, visual language llm, RAG, websearch, openwebui chatgpt alternative and tools calling and integration with local text editors and IDE)

Is this too much? What do folks want to see the most? Is there something folks get stuck on?

I want to do this for the people who know what they are doing. Think senior developers, devops, SRE, data scientists who know their way around linux and want to get things working from scratch locally. Article is ready-ish but I want to add more (or trim it down) based on feedback before posting.

EDIT: I want to also gear this article to those that have a single blackwell GPU with 32GB-96GB VRAM.

[ethicalTester]

Hey man, stumbled upon both of your posts while searching for a solution for the “CUDA unsupported” Fedora 43 that i just switched to from Linux Mint.

Would really appreciate such information as I’m already an AI engineering student and most of my work is involved around python and anything LLM/AI related.

I’d also love to know if you found some hacky way to solve the issue with the current glibc in fedora 43 which is causing NVIDIA not to support it (as far as i know) because it’s been a paintful experience to set up drivers let alone set up AI related libraries/tools.

I even considered going down to fedora 42 but fearing that it might also have unsopported newer dependencies.

Too long of a reply, but yes; such post will be very useful for AI folks. Looking forward for it

[progressEdd]

@level1mo

data science notebooks: setting up a jupyterlab and marimo to use with CUDA 13 including using uv for keeping everything together and latest pytorch with an example running

This would be good. Marimo makes development and quick UIs very nice. Used it for a project over the holidays recently

bedHedd's AI invasion

looks interesting, I had some time during the holidays and used it as a excuse to learn marimo for the front end [image] I think you should use structured outputs as you won’t need to write regex to parse out the premise, opening paragraph, etc. I have a example here from __future__ import annotations from typing import Optional, List from py…

Maybe a more advanced styling guide because mine is pretty bare bones

@level1mo

setting up a local ai stack (NVFP4/MXFP4 llm, visual language llm, RAG, websearch, openwebui chatgpt alternative and tools calling and integration with local text editors and IDE)

This would be good for either LM Studio (for beginners) and llama.cpp (for intermediate / advanced). I get frequent out of memory errors for lama.cpp with the qwen vision models. Open code would also be interesting too.

@level1mo

EDIT: I want to also gear this article to those that have a single blackwell GPU with 32GB-96GB VRAM.

If you do, you should enter it into this

The 4x AMD Radeon AI Pro R9700 128gb Giveaway

This is a companion topic to the video and AMD R9700 Giveaway. We built this MONSTER. It has 4 gpus. It needs a special case. And yet, it uses less power than a comparable system with 2x RTX 5090s (just barely). [We're Testing 4 Radeon AI Pro R9700s... What do you want to see? Community Event and Giveaway] … and I’m giving away two o…

As it would be good use case for the 9700s which are 32 gb of vram each.

If I were you, I’d link the previous post in the body of the new one as discourse will embed it

[ethicalTester]

After hours of research and stumbling upon the tutorial for setting up llama-cpp on fedora (it’s on their build guide, i can’t post links), I think i’ve discovered the best linux software ever; toolbx.

I don’t know how no one talked about this when discussing cuda issues on fedora, not even on the nvidia/fedora forums nor reddit.

Am i missing the bigger picture or is this the better solution instead of tinkering with dependencies directly on your system?